You are on page 1of 23

LQR and H

2
control problems
Domenico Prattichizzo
DII, University of Siena, Italy
MTNS - July 5-9, 2010
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 1 / 23
Geometric Control Theory for Linear Systems
Block 1: Foundations [10:30 - 12.30]:
Talk 1: Motivation and historical perspective, G. Marro [10:30 - 11:00]
Talk 2: Invariant subspaces, L. Ntogramatzidis [11:00 - 11:30]
Talk 3: Controlled invariance and invariant zeros, D. Prattichizzo [11:30 - 12:00]
Talk 4: Conditioned invariance and state observation, F. Morbidi [12:00 - 12:30]
Block 2: Problems and applications [15:30 - 17.30]:
Talk 5: Stabilization and self-bounded subspaces, L. Ntogramatzidis [15:30 - 16:00]
Talk 6: Disturbance decoupling problems, L. Ntogramatzidis [16:00 - 16:30]
Talk 7: LQR and H
2
control problems, D. Prattichizzo [16:30 - 17:00]
Talk 8: Spectral factorization and H
2
-model following, F. Morbidi [17:00 - 17:30]
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 2 / 23
Outline
Statement of the problem
Hamiltonian formulation
Geometric insight in the Hamiltonian system
Solving the cheap LQ problem in a geometric setting
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 3 / 23
LQR problems
Consider
x(t) = Ax(t) + B u(t) x(0) = x
0
y(t) = C x(t) + D u(t)
and the index to be minimized
J(x, u) =


0
y
T
(t) y(t)dt =


0
[
x
T
(t) u
T
(t)
]
[
C
T
C C
T
D
D
T
C D
T
D
] [
x(t)
u(t)
]
dt
=


0
(
x
T
(t) Q x(t) + 2 x
T
(t) S u(t) + u
T
(t) R u(t)
)
dt
with Q = C
T
C, S = C
T
D and R = D
T
D. Usually S = 0.
The problem is
Regular when R > 0 ker D = {0}
Singular when R 0 ker D {0}
Cheap when R = 0 ker D = IR
m
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 4 / 23
The main contributions
LQR in a geometric setting
The LQR problem for singular cases was investigated by Hautus,
Silvermann, Willems using distributions.
Saberi and Sannuti (1987) investigated the cheap (R = 0) and
singular (R 0) problems with the special coordinate basis.
Saberi, Sannuti and Stoorvogel (1992 and 1995) proposed a solution
based on LMIs.
Contribution
The geometric techniques are applied directly to the Hamiltonian system.
The cheap LQ problem admits solution with an algebraic feedback only if
the initial condition is contained in a special subspace.
This talk is based on Prattichizzo, Ntogramatzidis and G. Marro
(Automatica, 2008)
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 5 / 23
Statement of the problem - Cheap LQ problem
Consider the LTI system
x(t) = Ax(t) + B u(t) , x(0) = x
0
y(t) = C x(t)
with the assumptions:
(A1) (A, B) is stabilizable
(A2) has no invariant zeros on j IR
Determine the set of initial conditions x
0
such that a static state feedback
matrix K exists s.t.
AB K is stable
the corresponding state trajectory minimizes
J(x, u) =
1
2


0
y
T
(t) y(t)dt =
1
2


0
x
T
(t)C
T
C x(t) dt
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 6 / 23
Hamiltonian formulation
Consider the Hamiltonian function
H(t) :=
1
2
x
T
(t)C
T
C x(t) +
T
(t) (Ax(t) + B u(t))
and derive the state, costate equations and stationary condition as
x(t) =
(
H(t)
(t)
)
T
= Ax(t) + B u(t)

(t) =
(
H(t)
x(t)
)
T
= C
T
C x(t) A
T
(t)
0 =
(
H(t)
u(t)
)
T
= B
T
(t)
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 7 / 23
The Hamiltonian system
The optimal state trajectory and control law for our problem satisfy the
Hamiltonian system:
[
x(t)

(t)
]
=
[
A 0
C
T
C A
T
] [
x(t)
(t)
]
+
[
B
0
]
u(t)
y(t) =
[
0 B
T
]
[
x(t)
(t)
]
= 0
The matrices above are denoted by

A,

B and

C, while (

A,

B,

C) is
denoted by

.
Problem: nd u(t) = K x(t) such that for x(0) = x
0
y(t) =0 for all t 0
x(t) 0 for t .
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 8 / 23
Geometric approach: notations and properties
The geometric approach setting will require the following notations

: the maximum (A, imB)-controlled invariant subspace of


(A, B, C) contained in the null space of C.

: the minimum (A, kerC)-conditioned invariant subspace


containing the image of B.
F is a friend of

if
(A+B F)

At least one friend exists for every controlled invariant.


D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 9 / 23
Geometric approach
The Hamiltonian system is

x(t) =

A x(t) +

B u(t)
0 =

C x(t)
x =
[
x

]
The problem is equivalent to a disturbance decoupling problem in a
state-costate domain.
The cheap LQR has solutions i
an internally stabilizable output
nulling subspace of

exists whose
projection on the state space of
contains x
0
.
Classical DDP conditions

R
ker

C and
modes on

R
must be stable.
x
0

R
2n
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 10 / 23
Solving the cheap LQR problem
1 compute

;
2 compute a matrix

F such that (

A+

F)

3 compute

R
, the maximum internally stable (

A+

F)-invariant
contained in

;
4 if x
0
(

R
) the problem admits a solution F; if not, the problem
has no solution.
Next
What about the geometric structure and dimension of

R
?
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 11 / 23
Further notation
is left invertible i

={0}.
is right invertible i

=IR
n
.
Adjoint system
Let
T
= (A
T
, C
T
, B
T
) be the adjoint of = (A, B, C).
is left invertible i
T
is right invertible
is right invertible i
T
is left invertible
Time-reversed system

1
= (A, B, C) is the time-reversed system associated with .

1
and

1
is left invertible i
1
is left invertible
is right invertible i
1
is right invertible
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 12 / 23
A geometric insight in the Hamiltonian system

is the series connection of and


T






T
u(t)
v(t)
y(t)

left invertible =

is left and right invertible.
Left invertibility of

[IEEE TAC, 2002].
Right invertibility of

comes from equivalence of

T
and

1
through a coordinate transformation.
The set of invariant zeros of

consists of all the invariant zeros of

along with their opposite, hence they are all coupled by pairs (z, z).
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 13 / 23
How the second block is functionally driven






T
u(t)
v(t)
y(t)

Let U be a basis matrix of the subspace of functional controllability C

,
and denote by the symbol

the new system (A
T
, C
T
U, B
T
).



U
w(t) y(t)
v(t)

T
System

is left and right invertible, and

T
=

T
and
dim

= dim

= r
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 14 / 23
Relating geometric subspaces of and


T
u(t)
v(t)
y(t)

Denote by S and by

S two basis matrices for

and

, respectively.
The largest input containing subspace

of

has the following structure:

= im
[
S
0

S
]
and dim

= 2r
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 15 / 23
Main result
x
0

2n
r n r
If (A, B, C) is left invertible, then

is left
and right invertible and

= IR
2n
.
=
being dim

= 2r , it follows that
dim

= 2(n r ).
Modes in

are the zeros of


that are all coupled by pairs (z, z).


Modes in

are stable and unstable by pairs.


D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 16 / 23
Recall: Statement of the problem - Cheap LQ problem
Consider the LTI system
x(t) = Ax(t) + B u(t) , x(0) = x
0
y(t) = C x(t)
with the assumptions:
(A1) (A, B) is stabilizable
(A2) has no invariant zeros on j IR
Determine the set of initial conditions x
0
such that a static state feedback
matrix K exists s.t.
AB K is stable
the corresponding state trajectory minimizes
J(x, u) =


0
y
T
(t) y(t)dt =


0
x
T
(t)C
T
C x(t) dt
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 17 / 23
Recall: Disturbance decoupling
[
x(t)

(t)
]
=
[
A 0
C
T
C A
T
] [
x(t)
(t)
]
+
[
B
0
]
u(t)
y(t) =
[
0 B
T
]
[
x(t)
(t)
]
= 0
Problem: nd u(t) = K x(t) such that for x(0) = x
0
y(t) =0 for all t 0
x(t) 0 for t .
This can be achieved if and only if x
0
is such that
0
exists so as that
[
x
0

0
]
belongs to an internally stabilizable output-nulling subspace of

.
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 18 / 23
Resolvent subspace
x
0

2n
r n r
If (A, B, C) is left invertible

= IR
2n
Modes in

are stable and unstable by pairs.


The dimension of
R
to which x
0
must belong for the solvability of the
cheap problem is n r .
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 19 / 23
The main results
Let be left invertible. An (n r )-dimensional internally stabilizable
output nulling subspace

R
of

exists, such that all its n r poles,
all unassignable, are stable.
Let be left invertible. The cheap problem is solvable if and only if
x
0
(

R
) i.e., if and only if x
0
belongs to the projection of

R
on
the state-space of .
The static feedback
Let

F
R
a friend of

R
. Partition a basis matrix

V
R
of

R
and

F
R
as

V
R
=
[
V
X
V

]
,

F
R
=
[
F
X
F

]
the state-feedback matrix K solving the problem is given by
K = (F
X
+ F

V
+
X
)
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 20 / 23
Cheap vs. regular problems
The dimension of
R
is n r where r is the dimension of

.
Then
r = 0 and n r < n
there always exists x
0

R
for which the cheap LQ problem does not
admit solutions with static feedback.
Regular problems. If the problem is
regular, ker D = {0}, then dim

=
0 and there always is a solution for any
x
0
.
x
0

2n
r n r
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 21 / 23
Extension to non left invertible systems
Consider the auxiliary system (A+B F, B U, C), where
1. F is such that (A+B F)

, and all the eigenvalues of


(A+B F) restricted to

(which are free) are stable


2. U is a basis matrix of (B
1

The system thus obtained is left invertible.


Now, let K
F
be the optimal state-feedback matrix for the auxiliary system.
Matrix
K := U K
F
F
is one of the many solutions of the original problem.
Dierent solutions correspond to dierent choices of matrix F.
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 22 / 23
Concluding remarks
A new approach to the solution of the cheap linear quadratic
regulator problem has been presented for continuous-time systems.
This method is based on a geometric characterization of the structure
of the Hamiltonian system.
This insight is achieved by using the standard tools of the geometric
approach, without resorting to changes of basis and reduced order
algebraic Riccati equations.
By recasting the cheap LQ problem as a perfect decoupling problem
in the Hamiltonian system, we derive:
1 the dimension of the subspace of the admissible initial conditions
2 a basis matrix of this subspace, which is a controlled invariant solving a
standard decoupling problem with stability for the Hamiltonian system
3 an optimal state to input feedback matrix F, which is derived as a
straightforward function of a friend of the above controlled invariant
D. Prattichizzo (Siena, Italy) MTNS - July 5-9, 2010 23 / 23